MedSAM2-Anatomy boosts hip MRI segmentation to 0.92 Dice without retraining

Surgical planning for hip and shoulder procedures relies on high-resolution 3D segmentation of CT and MRI scans, but standard frozen AI models often fail when faced with new imaging domains. CNN-based expert models like TotalSegmentator are automatic yet inflexible, while promptable foundation models such as MedSAM2 generalize better but require manual prompting. MedSAM2-Anatomy, created by John Garcia Henao and 13 colleagues from ETH Zurich, Balgrist University Hospital, and other institutions, sidesteps this tradeoff entirely. It uses a frozen expert model to generate anatomical priors, automatically converts those priors into multiple prompt hypotheses for the frozen MedSAM2 foundation model, and fuses the resulting candidate masks while rejecting anatomically implausible outputs. No model weights are updated, and no human interaction is required.

Evaluated on the independent Balgrist-V0 cohorts, the framework delivered dramatic gains: median Dice on hip MRI improved from 0.71 to 0.92, and on shoulder CT from 0.89 to 0.92. The median Hausdorff distance (HD95) on hip MRI dropped from 22.0mm to 5.0mm, a fourfold improvement in boundary accuracy. Interestingly, on the public TotalSegmentator benchmark, the standalone expert model remained strongest—revealing that the optimal fusion strategy depends on how reliable the expert prior is. This insight suggests MedSAM2-Anatomy works best when expert priors are noisy or imperfect, making it a practical, plug-and-play solution for clinical deployment where retraining is costly and annotated data are scarce.